From: Nisha Gopalakrishnan
Date: June 6 2012 11:14am
Subject: bzr push into mysql-trunk branch (nisha.gopalakrishnan:3949 to 3950)
 Bug#11827404
List-Archive: http://lists.mysql.com/commits/144107
X-Bug: 11827404
Message-Id: <201206061114.q56BEtAj025514@acsmt356.oracle.com>
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: 7bit

 3950 Nisha Gopalakrishnan	2012-06-06
      Bug#11827404:60114: INSERT DELAYED CAN CALL HA_RELEASE_AUTO_INCREMENT
                          WITHOUT HAVING A LOCK.
      
      Analysis
      -------
      
      Mysqld crashes at times on a debug build, during the execution of
      the INSERT DELAYED statement under error scenarios. When a thread
      executes a DELAYED statement for a table, a handler thread is
      created to process all DELAYED statements for the table. The handler
      thread locks the table and handles the queued up inserts on the table.
      The handler thread updates the auto increment value row prior to
      writing the record. After the handler thread processes the stacked
      inserts, the auto increment value is reset and the table is unlocked.
      In case of simple INSERT, the thread that received INSERT request
      resets the auto increment value after writing the record. But not
      in the case of DELAYED INSERT since the threads share the table
      object.
      
      Under error scenarios, the thread that received the INSERT DELAYED
      request makes a redundant call to reset the auto increment value.
      
      The crash is observed in an error scenario and under a timing window
      where:
      The handler thread resets the auto increment value and unlocks
      the table, then the thread which received the INSERT DELAYED request
      tries to call the redundant release of auto increment value without
      having the lock on the table.
      
      Fix:
      ---
      Since the handler thread ensures the cleanup, removed the redundant
      call to reset the auto increment value in the main thread. Also in
      case of simple INSERT, it was observed that the call to reset auto
      increment value was redundant since the release was performed much
      ahead during the execution.
      
      NOTE:No regression test case has been added because the sync points
      currently does not work for the handler thread. The patch has
      been verified by running the existing test case main.delayed which
      fails sometimes without this patch.

    modified:
      sql/handler.cc
      sql/sql_insert.cc
 3949 Sunny Bains	2012-06-06
      BUG 14161424 - IMPORT OF COMPRESSED TABLE DOES AN EXTRA CHECKSUM OF NON INDEX PAGES
      
      When converting a page from the to be imported tablespace only checksum up to
      the compressed table page size. 
      
      Approved Marko Makela

    modified:
      storage/innobase/row/row0import.cc
=== modified file 'sql/handler.cc'
--- a/sql/handler.cc	2012-05-28 07:35:47 +0000
+++ b/sql/handler.cc	2012-06-06 11:13:22 +0000
@@ -3351,10 +3351,8 @@ void handler::get_auto_increment(ulonglo
 
 void handler::ha_release_auto_increment()
 {
-#ifdef BUG60114_IS_FIXED
   DBUG_ASSERT(table_share->tmp_table != NO_TMP_TABLE ||
               m_lock_type != F_UNLCK);
-#endif
   release_auto_increment();
   insert_id_for_cur_row= 0;
   auto_inc_interval_for_cur_row.replace(0, 0, 0);

=== modified file 'sql/sql_insert.cc'
--- a/sql/sql_insert.cc	2012-05-28 11:41:46 +0000
+++ b/sql/sql_insert.cc	2012-06-06 11:13:22 +0000
@@ -1097,8 +1097,6 @@ exit_without_my_ok:
   if (lock_type == TL_WRITE_DELAYED)
     end_delayed_insert(thd);
 #endif
-  if (table != NULL)
-    table->file->ha_release_auto_increment();
   if (!joins_freed)
     free_underlaid_joins(thd, &thd->lex->select_lex);
   thd->abort_on_warning= 0;

No bundle (reason: useless for push emails).